Study: AI assistants frequently make errors on news topics

The investigation found that errors ranged from mixing real news with parody to providing incorrect dates or even inventing events entirely. Out of more than 3,000 AI-generated answers, 45 percent contained at least one significant issue, regardless of language or country.

Major accuracy and sourcing problems

The report evaluated OpenAI’s ChatGPT, Microsoft’s Copilot, Google’s Gemini, and Perplexity. It revealed that one in five responses contained major inaccuracies, including hallucinated or outdated information.

Of the four tools tested, Google’s Gemini performed the worst, with serious issues in 76 percent of its answers—more than twice as many as the other assistants. The study attributed this to Gemini’s weak sourcing reliability and tendency to confuse satire with factual reporting.

Test conducted across 18 countries

Between late May and early June, 22 public broadcasters from 18 mainly European countries asked identical news questions to the AI assistants. The most common problem was outdated information.

For instance, when asked “Who is the Pope?”, several assistants—including ChatGPT and Copilot—incorrectly answered “Pope Francis”, even though, at the time, he had been reported deceased and succeeded by Pope Leo XIV.

In another example, Google’s Gemini falsely claimed that Elon Musk made a Nazi salute at Donald Trump’s inauguration, misinterpreting a satirical column as genuine news.

Experts warn AI tools are unreliable for news

“AI assistants are still not a reliable way to access and consume news,” said Jean Philip De Tender, deputy director general of the EBU, and Pete Archer, head of AI at the BBC.

Despite these shortcomings, the tools are increasingly used for information gathering, particularly among younger audiences. A Reuters Institute report published in June found that 15 percent of people under 25 now rely on AI assistants weekly for news summaries.